TL;DR

Google has released Gemma 4, its most capable open-source model family, in four sizes ranging from 2 billion to 31 billion parameters. The models ship under an Apache 2.0 licence — a shift from previous, more restrictive terms — and include native vision, audio input, and support for over 140 languages.

Four models for different hardware

The Gemma 4 family comprises four variants designed for specific deployment scenarios. The 31B Dense and 26B Mixture of Experts (MoE) models target workstations and servers, with Google claiming the 31B currently ranks third among open models on the Arena AI text leaderboard. The MoE variant activates only 3.8 billion of its total parameters during inference, trading some raw quality for faster response times.

At the smaller end, the E2B and E4B models are engineered for phones, IoT devices, and edge hardware. Google says these run offline on devices including Raspberry Pi and NVIDIA Jetson Orin Nano. Android developers can prototype with the models today through the AICore Developer Preview.

All four models support vision and image processing at variable resolutions. The smaller models add native audio input for speech recognition. Context windows range from 128K tokens on the edge models to 256K on the larger variants.

The Apache 2.0 shift

Previous Gemma releases carried custom licences with usage restrictions. Moving to Apache 2.0 removes those barriers, giving developers full commercial rights and the freedom to modify and redistribute without Google’s oversight. The change follows sustained feedback from the developer community and aligns Google’s open models more closely with genuinely open-source norms.

Google reports over 400 million downloads and 100,000 community-built variants across the Gemma family to date.

Looking forward

The release intensifies competition in the open-weight model space, where Meta’s Llama and Mistral have held strong positions. For UK developers and businesses, Gemma 4’s on-device capabilities and permissive licence lower the barrier to building AI products that keep data local — a consideration that matters as data sovereignty concerns continue to grow.